Unite segmentation method of uneven sparse road point cloud considering multi-sampling radius

Abstract

Road high precision mobile LiDAR measurement point cloud is a digital infrastructure in the fields of high precision map, automatic driving, High-precision automatic semantic segmentation of road point cloud is a key research direction at present. aiming at the problem that the semantic segmentation accuracy of existing deep learning networks is low for the uneven sparse point cloud measured by mobile LiDAR system, a deep learning method is proposed to divide point cloud data according to spatial location and considers the sampling point radius of regional groups. According to the spatial position of different objects, the method extracts the high-dimensional features of sampling points, and achieves the improvement of semantic segmentation accuracy of variable point cloud measured by high-speed mobile LiDAR system and carries out semantic segmentation experiment of The average test accuracy is 97.6%, and the mIOU reaches 0.82. The results show that compared with existing methods, the semantic results show that compared with the existing methods, the semantic segmentation accuracy of the proposed method is significantly improved for the uneven sparse road point cloud of mobile LiDAR system.